Bachelor's and Master's Theses or Research Internship / Bachelor- und Masterarbeiten oder Forschungspraxis

Benchmarking State-of-the-Art, Graph-based Machine Learning Solvers for Distribution Grid Power Flow

This topic uses a data-driven ML framework developed at TUM EMT to assess the generalization performance and tradeoffs of Graph Neural Network based AC Power Flow solvers in distribution grids. Generalization refers to the solver’s ability to maintain stable and accurate performance when applied in new contexts, i.e. unseen distribution grids. The student will re-implement advanced power flow solvers from literature (including attention- and physics-based GNNs) and perform a large-scale evaluation of the models, including their robustness.

Research method: Prototyping

Research question:

• Can we quantify the generalization potential of the state of the art GNN power flow solvers to determine which are the most robust for grid learning?

• What are the tradeoffs between model choices? Ex. Generalizability vs computational speed vs supervised/unsupervised.

• Can we establish a standard for future distribution grid solver techniques?

Possible approach:

• Literature review on state of the art GNN-based power flow.

• Adoption of open-source models or re-implementation of close-sourced models
• Evaluation and model generalization quantification using provided framework
• Evaluation of other model tradeoffs (i.e. computation, learning assumptions, sensitivity)
• Interpretation of practical implications of the results

Ehimare Okoyomon
e.okoyomon@tum.de

Generating Representative AC-OPF Datasets

Research method: Prototyping

Research question:

• How to systematically generate representative AC OPF datasets that span the feasible region?
• How to measure the generated data level of representativeness?

Possible approach: 

• Understand state-of-the-art based on literature
• Enhance existing state-of-the-art methods to increase level of representativeness
• Benchmark against dataset generated with SOTA approaches

Artificial Neural Networks for Optimal Power flow

Research method: Prototyping

Research question:

• What model design is suitable for improving ANNs optimal power flow predictions?
• What training approaches are suitable for maintain predictions feasibility and optimality?

Possible approach: 

• Understand state-of-the-art based on literature
• Implement ANN model for optimal power-flow prediction
• Benchmark results against SOTA ANNs

Spatiotemporal Assessment of Electric Vehicle Adoption Scenarios on the Distribution Grid

Research method: Prototyping/ simulation

Research question:

• What are the impacts of different electric vehicle (EV) penetration levels on the reliable operation of distribution grids?
• How can various EV adoption scenarios be effectively modelled to capture realistic spatial and temporal load variations?

Possible approach: 

• Understand state-of-the-art based on literature
• Model realistic grid demand incorporating different EV adoption rates
• Analyse results to quantify implications on grid operation

Your background/interests:

• Power-Flow analysis
• Programming experience - python

Arbel Yaniv

arbel.yaniv@tum.de

Benchmarking Active-learning Approaches for Optimal Power-Flow Dataset Generation

Research method: Prototyping

Research question:

• How to systematically generate representative AC OPF datasets that span the feasible region?
• What is the impact of the query approach on the model's predictive performance?

Possible approach: 

• Understand data sampling approaches based on literature
• Implement active learning pipeline (query strategy + ML model) to guide data sampling of OPF instances
• Benchmark various sampling approaches based on the scikit-activeml library 

Your background/interests:

• Python programming skills
• Machine learning
• Optimal power-flow

Resources:

[1] Herde, Marek, et al. "scikit-activeml: A Comprehensive and User-friendly Active Learning Library." (2025).

Non-Intrusive Load Monitoring with Active Learning

Research method: Prototyping

Research question:

• What is the impact of different acquisition functions on the performance of appliances disaggregation?
• What is the best fine-tuning approach, both layer-wise and sample-wise?

Possible approach: 

• Understand state-of-the-art based on literature
• Leverage scikit-activeml to analyse different active-learning approaches for NILM
• Try different fine-tuning experimental setups

Tanoni, Giulia, et al. "A weakly supervised active learning framework for non-intrusive load monitoring." Integrated Computer-Aided Engineering 32.1 (2025): 39-56.

Designing domain-specific priors for Prior-Fitted-Networks in Energy Management

Prior-Fitted-Networks [1] have shown intriguing success on small-scale tabular classification and regression tasks. Many energy management problems can be formulated as such classification or regression tasks. The TabPFN [2] uses a prior (=data generating mechanism) from either a Bayesian neural network or a structural causal model. These are designed for general tasks. The question is, can we achieve better performance by designing a prior that is more suited to a specific energy management task?

Research method: Modeling/ prototyping

Research question:

• Can Prior-Fitted-Networks trained on domain-specific priors outperform TabPFN and/ or traditional ML models?
• Which priors are suitable for <your selected energy management task>?

Possible approach: 

• Understand prior-fitted-networks and TabPFN from the literature
• Select energy management task (thermal building modeling, solar forecasting, wind forecasting, load forecasting, NILM)
• Suggest & test different priors for PFNs
• Benchmark results against TabPFN and a traditional ML model

Your background/interests:

• Interested in energy management and machine learning
• Statistics
• Programming experience - python

Resources:

[1] Müller, Samuel, et al. "Transformers can do bayesian inference." arXiv preprint arXiv:2112.10510 (2021). https://arxiv.org/pdf/2112.10510

[2] Hollmann, Noah, et al. "Tabpfn: A transformer that solves small tabular classification problems in a second." arXiv preprint arXiv:2207.01848 (2022). https://arxiv.org/pdf/2207.01848

Jan Marco Ruiz de Vargas

janmarco.ruiz@tum.de 

Improving ML models with pre-training from priors inspired by Prior-Fitted-Networks

Prior-Fitted-Networks [1] have shown intriguing success on small-scale tabular classification and regression tasks. Many energy management problems can be formulated as such classification or regression tasks. Can the priors used in PFN training also be useful for improving traditional ML model performance on energy management tasks?

Research method: Modeling

Research question:

• Does an ML model with access to the data from a PFN-prior achieve superior performance compared to one without?
• Which training strategies use this additional training data most optimally?

Possible approach: 

• Understand prior-fitted-networks and TabPFN from the literature
• Select energy management task (thermal building modeling, solar forecasting, wind forecasting, load forecasting, NILM) and ML model
• Suggest & test different pre-training strategies for the ML model
• Benchmark results against TabPFN and a traditional ML model

Your background/interests:

• Interested in energy management and machine learning
• Statistics
• Programming experience - python

Resources:

[1] Müller, Samuel, et al. "Transformers can do bayesian inference." arXiv preprint arXiv:2112.10510 (2021). https://arxiv.org/pdf/2112.10510

[2] Hollmann, Noah, et al. "Tabpfn: A transformer that solves small tabular classification problems in a second." arXiv preprint arXiv:2207.01848 (2022). https://arxiv.org/pdf/2207.01848

Jan Marco Ruiz de Vargas

janmarco.ruiz@tum.de 

Applying Prior-Fitted-Networks to Energy ML Tasks

Prior-Fitted-Networks [1] have shown intriguing success on small-scale tabular classification and regression tasks. Many energy management problems can be formulated as such classification or regression tasks. How good does a PFN then perform on energy ML tasks compared to other ML methods?

Research method: Benchmarking

Research question:

• For <energy ML task>, how good does TabPFN or its variants perform on <energy ML task> compared to other state of the art methods?
• Possible energy ML tasks (select one!):
  • Wind Power Forecasting (very short term = time series modeling)
  • Wind Power Forecasting (numerical weather prediction → power = power curve modeling)
  • Thermal building modeling
  • Solar Power Forecasting
  • Load Forecasting
  • Non-intrusive load monitoring

Possible approach: 

• Understand prior-fitted-networks and TabPFN from the literature
• Select energy management task (thermal building modeling, solar forecasting, wind forecasting, load forecasting, NILM) and SOTA benchmark ML models
• Benchmark results 

Your background/interests:

• Interested in energy management and machine learning
• Statistics
• Programming experience - python

Resources:

[1] Müller, Samuel, et al. "Transformers can do bayesian inference." arXiv preprint arXiv:2112.10510 (2021). https://arxiv.org/pdf/2112.10510

[2] Hollmann, Noah, et al. "Tabpfn: A transformer that solves small tabular classification problems in a second." arXiv preprint arXiv:2207.01848 (2022). https://arxiv.org/pdf/2207.01848

Jan Marco Ruiz de Vargas

janmarco.ruiz@tum.de 

Powering the Future of AI Energy Research: Building a Smart Data Loader for the e-SparX Platform

The application of Artificial Intelligence in the energy sector is crucial for tackling climate change, but progress is often hindered by a major bottleneck: accessing and operationalizing high-quality data. At the TUM Chair of Energy Management Technologies, we are developing 

e-SparX [1], a cutting-edge platform designed to accelerate Machine Learning research by making artifacts like datasets, models, and code transparently shareable and reusable.

To truly unlock the potential of our extensive TUM-EMT Open Energy Data Collection, we need to move beyond static tables and create a dynamic, on-demand data pipeline. This is where you come in. Your mission will be to design and build a cornerstone component for our ecosystem: a containerized, API-driven data loader. This tool will serve as the bridge between raw, open-source energy data (like high-resolution load and solar traces from datasets such as Pecan Street) and the AI models developed on the e-SparX platform.

This project isn't just about writing a script; it's about engineering a robust, scalable solution that will empower researchers to seamlessly pull the exact data they need, right when they need it, supercharging the pace of innovation in sustainable energy research.

Research method: Prototyping

Research question:

• How can we design and implement a robust, containerized data loader to programmatically serve datasets from our Open Energy Data Collection for direct use in AI research workflows on the e-SparX platform?

• What API design is most effective for researchers to query and retrieve specific time-series data (e.g., load profiles, solar traces) with high temporal resolution (e.g., up to 1 second)?

• How can this data loader be integrated with the e-SparX artifact registry to enable seamless data discovery and operationalization within ML pipelines?

Possible approach: 

• Familiarize yourself with the architecture of the e-SparX platform and the contents of the TUM-EMT Open Energy Data Collection.

• Design the architecture for the data loader, including the API specifications (e.g., using FastAPI) and a suitable data handling backend.

• Implement the core functionality to ingest, process, and serve key energy datasets (e.g., Pecan Street), with a focus on high-resolution time-series data.

• Containerize the entire data loader service using Docker, ensuring portability and ease of deployment.

• Integrate the loader with our internal e-SparX version, enabling users to programmatically pull data and register it as new, discoverable artifacts within their research pipelines.

• Demonstrate the success of your solution with a clear use case, such as training a simple forecasting model using data served directly from your loader.

What we offer:

• Close supervision and deep integration into our motivated research team.

• Access to the e-SparX platform and infrastructure.

• The opportunity to build a critical open-source tool that will be used by the wider energy research community.

• Hands-on experience with a modern MLOps and Data Engineering tech stack (Python, FastAPI, Docker).

Your background/interests:

• Strong interest in Machine Learning, AI, and their application in the sustainable energy sector.

• Excellent programming skills in Python.

• Experience (or a strong desire to learn) with modern software engineering concepts, including building APIs (e.g., FastAPI, Flask), working with databases, and containerization (Docker).

Resources:

[1] Schneider, Annika, et al. "e-SparX: A Graph-Based Artifact Exchange Platform to Accelerate Machine Learning Research in the Energy Systems Community." Proceedings of the 16th ACM International Conference on Future and Sustainable Energy Systems. 2025.

Manuel Katholnigg

manuel.katholnigg@tum.de 

Finetuning Multi-Modal LLMs on Energy ML Tasks

Colleagues have fine-tuned MMLLMs on a tabular classification task [1], with significant improvement in prediction accuracy. This is a puzzling result, why does this work? Can it work with other prediction tasks? In this topic, we will investigate this by applying LLM finetuning to energy ML tasks. 

Research method: Benchmarking

Research question:

• For <energy ML task>, how good does a fine-tuned MMLLM perform on <energy ML task> compared to other state of the art methods?
• Possible energy ML tasks (select one!):
  • Wind Power Forecasting (very short term = time series modeling)
  • Wind Power Forecasting (numerical weather prediction → power = power curve modeling)
  • Thermal building modeling
  • Solar Power Forecasting
  • Load Forecasting
  • Non-intrusive load monitoring

Possible approach: 

• Understand how to finetune LLMs and how to properly feed it the relevant data. 
• Select energy management task (thermal building modeling, solar forecasting, wind forecasting, load forecasting, NILM) and SOTA benchmark ML models
• Benchmark results 

Your background/interests:

• Interested in energy management and machine learning
• Statistics
• Programming experience - python

Resources:

[1] Domiter, Andrea, and Srinivasan Keshav. "Machine Learning for Building-Level Heat Risk Mapping." Proceedings of the 16th ACM International Conference on Future and Sustainable Energy Systems. 2025.

Jan Marco Ruiz de Vargas

janmarco.ruiz@tum.de 

Adaptive Control of Electric Bus Depots for Grid Services

Keywords: Learning methods, electric buses, grid flexibility

Objective:
Electric bus depots are considered as passive grid burdens. The objective of this MSc thesis will be to investigate how electric bus depots can be transformed into flexible assets deploying advanced control strategies for energy management.

Research questions:

• How real-time fleet management strategies improve the reliability of electric bus operations under practical system constraints?
  
• How do different real-time control approaches compare in managing disruptions in deadline-constrained fleet operations?

Expected background:

• Interested in energy management and machine learning
• Mathematical modeling
• Programming experience (Python)

Biswarup Mukherjee

biswarup.mukherjee@tum.de